Agentic AI in Capital Markets
Agentic AI becomes meaningful in capital markets when systems move beyond assisting with tasks to taking decisions and actions inside governed financial workflows. The opportunity is real, and so is the constraint: an agent that acts on a client's behalf inherits every governance requirement a person would, and several new ones.
From assistant to actor
The first wave of generative AI in banking was assistance: a chat window, a summary, a draft. Agentic AI changes the unit of work. Instead of answering a question, the system carries out a sequence: it pulls market and proprietary data, reasons over it, drafts the output and hands it to a banker for judgement. That is where the productivity gain compounds, and where the risk compounds too.
Where it works
The strongest early uses share three features: several data sources, repeated structure, and a human who makes the final call. Examples include meeting preparation, investor and target screening, precedent research across transaction documents, and first drafts of client materials. The shared idea is the augmented banker: AI in every workflow, built on the firm’s own data rather than on whichever model is in fashion.
What has to be true first
- Proprietary data, connected safely. Agents are only as useful as the data they can reach. Orchestration layers such as MCP connect agents to internal and market data while controlling access and keeping data inside the firm.
- Traceable steps. Every action and source should be logged, so a reviewer can see how an output was produced.
- Human sign-off at the right point. Not on every step, which destroys the gain, but where a decision reaches a client or the market.
- Measured reliability. Multi-step systems fail in multi-step ways. Reliability has to be measured per workflow, not assumed from a model’s benchmark.
The economics are set by review burden. As The Checking Problem shows, an AI workflow’s value depends less on how often it is right than on how much of it a human must still check.
In practice
Prerit Ahuja is building agentic, MCP-orchestrated AI into the workflows of DNB Carnegie’s Investment Banking Division, connecting AI to proprietary and market data under a governed platform. He spoke on autonomous operations in finance at the Agentic AI & Automation in Finance Summit in Stockholm in 2026.
Frequently asked questions
What is agentic AI in capital markets?
AI systems that plan and carry out multi-step work, such as gathering data from several sources, drafting analysis and preparing materials, rather than answering a single prompt. In capital markets this spans origination, execution support, investor targeting and research.
Where does agentic AI create value in investment banking?
In workflows that combine several data sources and repeated judgement: preparing for client meetings, screening targets or investors, assembling precedents, and producing first drafts of materials so bankers start from most of the way there rather than from zero.
What are the risks of agentic AI in regulated finance?
Compounding errors across steps, actions taken on unverified data, unclear accountability, and review burden that grows faster than the work saved. Each step needs traceable sources and a clear point of human sign-off.
How do firms connect AI agents to proprietary data safely?
Through a governed orchestration layer, for example Model Context Protocol (MCP) connectors, that controls which data and tools each agent may reach, logs what it did, and keeps proprietary data inside the firm's boundary.